Convolutional neural networks (CNN)–based liver segmentation on CT and MRI medical images
Loading...
Files
Date
Authors
Λίβανος, Γεώργιος
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Medical Image is important in different aspects of many medical diagnostic, treatment and follow-up procedures. Especially images from modalities like Computed Tomography (CT), Magnetic Resonance Imaging (MRI) and Ultrasound (US), due to their minimal to non-invasive nature, provide essential diagnostic information of the internal anatomy of the patient.In the present thesis we investigate automatic healthy liver segmentation from Computed Tomography and Magnetic Resonance Imaging scans by using convolu-tional neural networks. More specifically, based on the U-net architecture, we investigate some variations which include adding batch normalization layers, dropout layers and using leaky-ReLU activation function. As a loss function we utilize a combination of a weighted cross-entropy function and a weighted variation of generalized dice loss function. We use this setup to train three models with CT only, MRI only and both type of scans, respectively. We also apply three data augmentation techniques.To test the proposed methodology, we evaluate the performance of all different variations of the U-net, training these models on the raw data. Subsequently, we apply different augmentation strategies to the best performing model of the first stage for each of the three data categories.
Description
Keywords
Machine learning, Deep learning, Convolutional neural networks, Liver segmentation, Unet